Use your ChatGPT subscription¶
An administrator can connect their own ChatGPT Plus or Pro subscription and use its models in Numi without separate API billing.
The connection is strictly personal to the admin who signs in:
- Only that admin sees the
chatgpt:<slug>models in the model selector and can run turns with them. - Other users and other admins never see or access the connection, and access tiers cannot grant it.
- Background jobs (chat titles, memory extraction, and tool routing) continue to use regular catalog models.
Numi connects through OpenAI's documented
Sign in with ChatGPT token sharing
flow and sends requests to the standard Responses API
(https://api.openai.com/v1/responses). It does not use the Codex backend or
Codex CLI tokens.
Connect¶
- Open Admin and locate the ChatGPT subscription card.
- Click Connect. A ChatGPT sign-in tab opens; sign in and grant permission.
- ChatGPT redirects to
http://127.0.0.1:4567/auth/callback: - If your browser runs on the same machine as Numi (or forwards port 4567 locally), the callback completes automatically and displays ChatGPT connected.
- If Numi runs on a remote server and
127.0.0.1:4567does not load in your browser, copy the full URL from the browser's address bar, paste it into the Admin card, and click Finish.
Sign-in attempts expire after CHATGPT_SIGN_IN_TIMEOUT_MINUTES (default 10
minutes) and are single-use.
Choose models and settings¶
Once connected, the card lists every model available on your plan with its supported reasoning levels and context window. All models start enabled; toggle off any you do not want in the chat picker. Use Refresh models to re-fetch the plan's current model list.
Four plan-wide settings appear below the model list:
| Setting | Request field | Default | Effect |
|---|---|---|---|
| Answer length | text.verbosity: low, medium, high |
Medium | Controls how detailed responses are |
| Reasoning summary | reasoning.summary |
Auto | Controls how much reasoning summary is streamed to the chat |
| Fast mode | service_tier: "priority" |
Off | Requests priority processing (consumes plan allowance faster) |
| Full-resolution images | detail: "original" |
Off | Sends images at original resolution on models that support it |
Reasoning summary options:
| Choice | Request value | Shown in chat |
|---|---|---|
| Auto summary (default) | auto |
Summary length chosen by ChatGPT |
| Short summary | concise |
Brief summary |
| Detailed summary | detailed |
Fullest summary provided by ChatGPT |
| No summary | omitted | No reasoning text shown |
How requests are sent¶
- Every request streams with
store: false, sending the system prompt as adevelopermessage along with the conversation history and attachments. - Each chat uses its own
prompt_cache_keyandsession_idheaders so follow-up turns can reuse the cached prompt prefix. - Numi requests
reasoning.encrypted_contentand replays those encrypted reasoning items (without item IDs) on subsequent turns to the same model so multi-step tool use and follow-up turns preserve reasoning state. Encrypted items are never sent to other models or included in share snapshots. - Token counts, cache hits, and reasoning tokens are recorded on each reply.
Your subscription covers the usage, so nothing is billed per request. To keep
costs comparable with your other models, each reply shows what the same
request would have cost through the OpenAI API, marked
≈as an estimate. - The API prices come from OpenRouter's public model list. Numi reads it when
you connect or refresh the model list (Refresh models), without a
credential and without sending anything about you or your chats. A model the
list does not know, or a connection made before this lookup existed, shows
$0until the next refresh finds a price. The Fast tier is estimated at the standard price.
Limits and troubleshooting¶
- Plan usage limits: When your ChatGPT plan hits a rate or usage cap, the error is shown in chat. Numi does not silently fall back to a paid API key.
- Disconnect: Clicking Disconnect deletes the local tokens and asks OpenAI to revoke them. Your model selection and plan settings are remembered if you reconnect later.
- Encryption key: Tokens are encrypted at rest with a key derived from
AUTH_SECRET_KEY. RotatingAUTH_SECRET_KEYinvalidates the saved connection and requires signing in again. - Single-process requirement: OAuth refresh tokens are single-use. Run Numi as a single server process per database so concurrent workers do not race on token refresh.
Configuration reference: CHATGPT_*.